Simulating Foot-and-Mouth Disease in the United States using the Animal Disease Spread Model

Transcription

1 Simulating Foot-and-Mouth Disease in the United States using the Animal Disease Spread Model Melissa Schoenbaum, Dawit Assefa, Lindsey Holmstrom, Amy Delgado USDA APHIS VS STAS CEAH Fort Collins, Colorado Josiah Seaman Newline Technologies Colorado Springs, Colorado ABSTRACT Animal disease outbreaks are a global concern. These outbreaks can cause food scarcity and economic losses, impact animal health, and, if zoonotic, affect human health. For an outbreak of a disease not currently found within a country, these consequences can be greater. Disease modeling is one tool that can be used to better understand disease spread and evaluate different control strategies. USDA APHIS VS CEAH addresses questions related to minimizing the consequences of outbreaks while protecting our nation s livestock, food supply, and the environment. Using modeling tools, analysts evaluate alternative control strategies, provide recommendations on planning and effective use of resources, develop surveillance strategies, and estimate the economic impacts of disease control options. In this paper, we used the Animal Disease Spread Model (ADSM) to simulate an outbreak of foot-and-mouth disease when introduced from multiple locations on the same day in a seven-state area of the United States. ADSM is a stochastic, spatially explicit compartmental model that simulates herd-to-herd spread of a highly contagious disease. The outbreak scattered in a variety of directions to other herds; however, only a few locations were infected consistently in a high percentage of iterations. Identifying locations at higher risk for disease introduction and/or spread can help decisionmakers target risk-based surveillance and control strategies. Since ADSM is a new model available to the global modeling community, this paper also discusses the modeling process required by analysts to parameterize and use the model and highlights how this simulation method can provide useful information for decision-makers. ABOUT THE AUTHORS Melissa Schoenbaum is a biological scientist with the USDA APHIS VS CEAH Monitoring and Modeling unit. She received her BS in biological sciences from East Texas State University and her MS in computer information technology from Regis University. She serves as the modeling team s data management specialist. Dawit Assefa is a Program Analyst with the USDA APHIS VS CEAH Monitoring and Modeling unit. He received a BA in economics and an MS in computer information system from Colorado State University, with an emphasis on business intelligence. He performs data visualization, validation, and cleaning functions using a multitude of software. Josiah Seaman is a software developer working in bioinformatics with a specialty in data visualization. He has a BS from Colorado State University in computer science and is working on a PhD in genetics at Kew Royal Botanical Gardens. Josiah designed a genome visualization program called DNASkittle and founded Newline Technical Innovations. He works with others at Newline on science and automation tools. Lindsey Holmstrom is a veterinary epidemiologist with the USDA APHIS VS CEAH Monitoring and Modeling unit. She received her BS in economics from Baylor University, her doctorate of veterinary medicine (DVM) degree from the College of Veterinary Medicine at Texas A&M University, and her PhD in veterinary epidemiology from the University of California, Davis (awarded March 2017). As part of the modeling team, she helps advance and apply epidemiologic models to develop scenarios involving disease spread, alternative control strategies, and associated impacts. Amy Delgado currently serves as the director of the Monitoring and Modeling Unit of USDA APHIS VS CEAH in Fort Collins, CO. Her work focuses on providing scientific support for emergency response planning for transboundary animal diseases. Dr. Delgado received a BS in animal science prior to receiving her DVM and PhD in epidemiology from Texas A&M University Paper No. 9 Page 1 of 10

2 Simulating Foot-and-Mouth Disease in the United States using the Animal Disease Spread Model Melissa Schoenbaum, Dawit Assefa, Lindsey Holmstrom, Amy Delgado USDA APHIS VS STAS CEAH Fort Collins, Colorado Josiah Seaman Newline Technologies Colorado Springs, Colorado SIMULATING ANIMAL DISEASE OUTBREAKS When a disease is absent from a population, modeling offers a useful method to simulate the spread of disease and evaluate control measures to manage an outbreak. Foot-and-mouth disease (FMD) is an economically important disease due to anticipated response program costs that can range from millions up to a billion dollars and market effects, namely trade embargoes on meat and livestock by international trade partners, that can bring anticipated economic losses into the billions of dollars (Thompson et al., 2002; Carpenter et al., 2011). The effects of an event would be distributed among many parties domestically including producers, consumers, service providers, and the government. Generally, trade bans are expected to be a major driver of total economic loss associated with FMD (Paarlberg et al., 2003). A confirmed outbreak would trigger border closures for United States exports, some lasting until such time as disease-free status is regained, although bilateral trade relations would determine the actual speed of markets reopening. For a country that sells a great deal of meat and livestock to international markets, the economic damage from FMD can be marked. The United States last experience with FMD was in 1929, leaving little available data to reference in dealing with an outbreak. Traditional emergency response has focused on strict movement restrictions and the rapid slaughter of infected and exposed livestock (Gibbens et al., 2001; McLaws and Ribble, 2007). In the case of FMD, the application of vaccination, with or without the subsequent destruction of vaccinates, has also been applied or explored in many countries (Bates et al., 2003a; Perez et al., 2004; Kitching et al., 2007; Barasa et al., 2008; Estrada et al., 2008). Emergency response exercises have recognized issues associated with a traditional stamping-out response to FMD in large feedlots, as depopulating and disposing of large numbers of animals is logistically and environmentally challenging. Alternative methods are needed for minimizing disease spread while allowing animals to reach their intended purpose. As such, modeling applications provide useful tools to decision-makers to evaluate different control strategies and plan for response before an outbreak occurs. Several models have been applied at state and regional levels to evaluate FMD outbreaks in the United States. The Davis Animal Disease Simulation (DADS) developed at the University of California, Davis was used to evaluate stamping-out, slaughter, and vaccination control strategies for FMD outbreaks in California and resulting economic costs (Bates et al., 2003a; Bates et al., 2003b; Carpenter et al., 2011). Similarly, the InterSpread Plus model, developed by Massey University in New Zealand, was applied in central California to identify disease and control parameters that have the greatest effect on FMD spread and evaluate delays in reporting, livestock movement, surveillance, tracing, vaccination, and depopulation on outbreak size and scope (Alok, 2013). The AusSpread model developed in Australia was used to investigate the impact of early detection, vaccine availability, and surveillance on control strategies from an FMD outbreak in the Panhandle region of Texas (Ward et al., 2009). The North American Disease Spread Model (NAADSM) has been used in many instances to simulate FMD outbreaks and evaluate control strategies and related economic costs on state and regional scales in the United States (Paarlberg et al, 2008; McReynolds et al., 2014; Miller et al., 2015; Pendell et al., 2015) Paper No. 9 Page 2 of 10

3 The U.S. Department of Agriculture, Animal and Plant Health Inspection Service, Veterinary Services, Center for Epidemiology and Animal Health (USDA APHIS VS CEAH) also addresses questions related to minimizing the consequences of animal disease outbreaks while protecting our nation s livestock, food supply, and the environment. Using a variety of modeling tools, analysts evaluate alternative control strategies, provide recommendations on planning and effective use of resources, develop surveillance strategies, and estimate the economic impacts of disease control options. Many of the modeling tools available, including those above, require continued development and enhancements as disease spread methods are better understood and control strategies change. Furthermore, there is a need for disease models with a user-friendly interface to support their use by analysts across the globe whose training in modeling may vary. As such, one of the current modeling initiatives of CEAH is developing the Animal Disease Spread Model (ADSM). ADSM is a stochastic, spatially explicit compartmental software application that simulates herd-to-herd spread of a highly contagious animal disease. Models such as ADSM allow decision-makers to compare strategies based on their costs and effectiveness in controlling disease. The recently developed ADSM is based on NAADSM, which was developed as a collaboration between USDA, the Canadian Food Inspection Agency, the University of Guelph, and Colorado State University (Harvey et al., 2007). At this time, ADSM is in beta testing with an expected release in fall It will be publically available at no cost. ADSM shares the same code with NAADSM but with several added functionalities. Like NAADSM, ADSM is flexible enough to simulate a variety of disease conditions, with an emphasis on highly contagious diseases that spread through some form of contact. The user can apply a variety of control measures, such as vaccination or depopulation of infected farms, to influence the outbreak. For the purposes of this paper, we used a previously developed NAADSM scenario, implemented in ADSM, in order to demonstrate a mature proof-of-concept in the new application. Since this is a new model that has not yet been detailed in the literature, this paper presents a description of the application and an example of the use of ADSM to evaluate a control strategy for a severe, hypothetical outbreak of FMD in the United States. OVERVIEW OF ADSM TECHNOLOGY The ADSM software architecture is divided into several components. The core simulation logic is referred to as the C Engine. The C Engine has been supporting NAADSM since its release in 2006 and has been exercised extensively by researchers around the world. This component is utilized in the shared layer of code between ADSM and NAADSM. It is written in C and highly optimized for running large-batch simulations. As implemented in ADSM, the C Engine takes as input a SQLite database file containing all the parameters for the scenario. As it runs, the simulation adds outputs to the database file, ensuring results are never separated from the parameters that created them. NAADSM scenarios import as XML files into ADSM, allowing projects to work on the new platform. The front-end user interface is built from the ground up using web technology to run on all platforms. The interface was built in Django and Python and presented as a web application in a browser window. When the user has an Internet connection, a linked glossary defining parameter inputs is accessible. The interface provides visual cues as users parameterize the scenario. Parameters are broken up by topic: population, disease, and controls. Details of each section are also divided into reusable pieces. A major difference from the earlier NAADSM user interface is that users now create a block of parameters, named with a context-appropriate name, and then assign where those parameters apply. Users can run one or many iterations of a scenario. While a batch of iterations is running, the interface provides live statistics on the current results. Results can be accessed in several different ways. Immediately upon completion, users can view and browse detailed outputs from the user interface, showing the distribution of all tracked variables across all iterations. SQL queries can be used to access the raw data from the tables in the SQLite database. Users have the ability to select several supplemental outputs to create herd-level event details for every day in every iteration. These.csv files are not created automatically due to the volume of data they would generate. ADSM user interface is built on web technology to enable easy maintenance and upgrades in the future as well as opening up a number of new usage options. The ADSM server is compiled for Windows and Linux machines. The server can be run locally on a single computer or hosted on a supercomputer on the same LAN. ADSM uses multithreading to maximize system utilization and get results back promptly. The Chrome browser can be used on any desktop computer to run simulations on the server. For example, an epidemiologist may use a Windows machine to run simulations on a Linux supercomputer without any issue Paper No. 9 Page 3 of 10

4 Currently anyone with access to a specific server shares a single file store. User logins and permissions systems are planned for a future release of ADSM. The current architecture is poised to be used as a cloud-based service after user separation is implemented. This would enable a central parameter library, contributed to by experts in their subject areas, which could be quickly queried in order to assemble scenarios. For studies of how parameters affect outcomes, variants can be built from existing scenarios. USING ADSM TO MODEL FMD The example presented in this paper is based on a previous analysis which used NAADSM in combination with the Paarlberg model (Paarlberg et al., 2008) to estimate economic impacts of disease (Delgado et al., 2015) under a variety of strategies. The spread and control of FMD within a seven-state region of the United States was simulated for a variety of strategies. For the purposes of this paper, a single scenario was selected and imported into ADSM. A synthetic population file was generated using the Farm Location and Animal Population Simulator (FLAPS) program, which disaggregated Census of Agriculture data from Texas and surrounding states livestock operations to simulate their locations (FLAPS, 2013). For the purposes of this example, the starting location of the outbreak was changed so that disease could be introduced in multiple locations on the same day in order to provide output data from large outbreaks. This starting setup demonstrates ADSM s capabilities for evaluating control options. The resulting simulated outbreaks were often severe, with impacts on the agricultural and broader economy. The remainder of the parameters used are consistent with those used by Delgado et al. (2015), which characterized the disease progression, the geospatial distribution of the susceptible population, the spatiotemporal interactions of farms within the population, and the control measures applied to the different farm types. OVERVIEW OF ADSM ADSM manages the three components necessary to simulate an animal disease outbreak: the susceptible population and their interactions, the progression of disease and subsequent transmission, and options for disease detection and control. First, the group of domestic livestock or poultry farms that are susceptible to the specific disease must be organized. Called a population in ADSM, these data are provided as a file, including the species and management practice for the animals, the count of animals in the herd/flock, and their geographic coordinates. The application uses production types to define both a species and a purpose for groups of animals, as differences in management practices affect disease spread. This example uses 12 productions types. These production types include a variety of species, management practices, and farm sizes since FMD is a disease that is contagious across multiple species. Second, the application uses a set of parameters defining disease progression and transmission. These inputs often vary by production type. Third, the application can simulate control activities, organized into protocols that address the specific production types. Parameters may include movement restrictions, contact tracing, zone-based surveillance, depopulation of infected herds/flocks, and vaccination of specified herds/flocks Paper No. 9 Page 4 of 10

5 Population The population of animals used in this example were from a simulated dataset representing a seven-state region within the United States as shown in figure 1. The full population file contained 363,989 farms of the 12 production types. The farms can be loaded into ADSM in the format exported from NAADSM (XML) or loaded as a comma-separated value file. Within the application, the population is visualized on a map and presented in tabular form. Filters and sorting functions are available, which allow users to navigate and explore the population data. Edits to individual farms, such as changing the disease state, can be applied from the population module of the ADSM application. Figure 1. Population Module Disease The disease portion of the application is flexible, allowing ADSM to represent several different highly contagious animal diseases. ADSM is a transition state model, meaning farms move between disease states. The disease states include susceptible, latent infection, subclinical infection, and clinical infection. The farm can also be in an immune state, either naturally immune or vaccine-induced immune. Parameters describing the disease states, such as how long a farm may be in a latent disease state, are represented using probability density functions in order to capture the inherent variability in biological processes. It is necessary to set rules for the model to determine the way the disease is spread between farms. These parameters describe the probability that disease moves between farms of a given production type that share a contact method. It is important that these parameters are specific to production types, as some species are likely to contact each other, while others are not. ADSM allows these parameters to be built into blocks which are given a user-defined, easily readable name, such as dairy to dairy indirect contact. This parameter block is then assigned to a source production type and a potentially exposed production type for this example, dairy cattle indirectly contacting other dairy cattle. A regular dairy business process is the milk truck that follows a route between dairies to pick up milk and deliver it to a processor, providing a mechanism for disease spread Paper No. 9 Page 5 of 10

6 Recall that FMD is a disease that spreads between multiple species; therefore, all 12 production types have the possibility to spread disease to the other production types. After assignment, we display a 12 production type by 12 production type by 3 contact methods matrix. The complexity of disease transmission becomes apparent as there are 432 possible options that could have unique spread parameters. It is likely that some production types share the same parameters, showing the value of the reusable parameter block. It is also important to note that some combinations do not have a contact method. Visualizing the matrix of disease spread allow users to quickly assess the assignments made in the previous step and identify gaps shown in white. The Review Disease Spread Module is shown in figure 2. Figure 2. Review Disease Spread Module 2017 Paper No. 9 Page 6 of 10

7 Controls ADSM provides multiple control measures that are generally used in managing animal disease outbreaks. As with the disease parameters, control parameters are built in blocks with a user-defined name. Each parameter block is then assigned to one or more production types. These controls describe how disease is detected in the absence of an outbreak and how surveillance and tracing activities are carried out once the outbreak response process starts. The controls can also account for imperfect diagnostic tests, as well as clinical examination of the animals by an owner or veterinarian. Farms may be depopulated in order to control the disease, and parameter settings define the time required for depopulation activities to be completed for each production type. In a large outbreak, delays in control activities are possible, resulting in prioritized waitlists for farms that need activities carried out. Radial, zone-based control measures are also available, and control options in combination with zone enforcement can be applied at the production-type level. Movement controls can be implemented to stop all or specific types of contact between production types. Figure 3 shows the Control Protocol Module with one control measure selected. Figure 3. Example of a Control Protocol with a relational function IMPLEMENTATION AND EXECUTION OF THE FMD MODEL In this example, a single scenario was exported from NAADSM and imported into ADSM. The small changes to the starting locations were made (see above), and 500 iterations of the scenario were run. The complete scenario run was completed in 13 hours, with approximately 100 seconds per iteration on a laptop computer. The summary statistics of the 500 iterations are shown in table 1. Outbreak size ranged from 3 to 86 herds, with an outbreak duration of 24 to 49 days. The number of animals depopulated ranged from 563,189 to 1,722,720. This range in outcomes reflects the stochastic nature of the application and the inherent uncertainty of complex systems. A full dataset is available in the sqlite3 database supporting the application, which can allow users to run statistical summaries and comparisons between scenarios when a variety of strategies are planned for analysis Paper No. 9 Page 7 of 10

8 Table 1 Summary Statistics Infected Farms 25 Median Value Infected Animals 338,100 Infected Farms at First Detection 13 Infected Animals at First Detection 199,020 Depopulated Farms 34 Depopulated Animals 946,403 Duration of Outbreak in days 31 Day of First Detection 6 Total Control Zone Area (km2) 25,418 Total Surveillance Area (km2) 61,355 Number of Control Zones 67 Number of Surveillance Zones 41 Immediate feedback for the user is available from within the application, providing a high-level overview of the scenario. Dynamic feedback is given as iterations are completed, and once all iterations have finished, a summary map is presented in figure 4. A subset of select, summarized variables is displayed, such as the median number of infected units and the median outbreak duration across all iterations. Figure 4. Map Output and Summary Results Users may also drill in to individual variables by navigating down the left menu shown in figure 4. The number of variables presented and graphed individually is dependent on how the scenario was parameterized. They can be viewed 2017 Paper No. 9 Page 8 of 10

9 as single iteration values or summarized across all iterations, depending on the level of detail needed by the user. This single iteration view can be valuable in exploring an extreme or unexpected simulated outbreak. Figure 5 shows the graph within ADSM from one iteration, by production type, with the cumulative number of farm units that were destroyed per day over the duration of the outbreak. Figure 5. Graph of Destruction of Units for a single iteration CONCLUSION In this scenario, disease was introduced on multiple locations on the same day. One of the most interesting results that this exercise produced was an outbreak that scattered in a variety of directions from the sources over the 500 iterations. In the map of results in figure 4 there are only a few locations that are hit consistently in a high percentage of iterations, shown in the darkest blue. Identifying locations at higher risk for disease introduction and/or spread can help decisionmakers target risk-based surveillance and control strategies to these areas. The starting locations would be included in these dark blue circles. In contrast, the farm dot in a white circle mean that it was included in at least a small percentage of iterations. The white circles range to all extents of the map. The supplemental output daily events could have provided the exact counts of animals involved in any iteration and those not included. ADSM is a flexible and powerful tool to simulate complex disease outbreak events. As shown in this example, the application can easily import and run scenarios from NAADSM, with reasonable run times, dynamic output visualization, and complex disease control options to explore. Users have access to a large, rich dataset of outputs for further analysis and comparison, allowing for the development and examination of many what if scenarios. In addition, new approaches that allow users to develop and reuse parameter blocks increase efficiency and allow for easier sharing of parameters between projects. The format of the user interface of ADSM makes it a good choice for beginning users, while the complexity of disease spread and control options makes it relevant for animal health professionals who need disease spread modeling support for decision-making. Work on ADSM continues USDA APHIS CEAH s commitment to the NAADSM collaboration with the aim of taking it into the future with improved and more accessible animal disease spread modeling applications Paper No. 9 Page 9 of 10

10 REFERENCES Alok, D. (2013). A high explanatory power model of foot and mouth disease spread in central California. Master s Thesis. Naval Postgraduate School. Monterey, CA: 107. Available at: Bates, T.W., Carpenter, T.E., Thurmond, M.C. (2003a). Benefit-cost analysis of vaccination and preemptive slaughter as a means of eradicating foot-and-mouth disease. Am J Vet Res, 64 (7), Bates, T. W., Carpenter, T.E., Thurmond, M.C. (2003b). Results of epidemic simulation modeling to evaluate strategies to control an outbreak of foot-and-mouth disease. Am J Vet Res, 64(2), Carpenter, T.E., O'Brien, J.M., Hagerman, A.D., McCarl, B.A. (2011). Epidemic and economic impacts of delayed detection of foot-and-mouth disease: A case study of a simulated outbreak in California. J Vet Diagn Invest 23, Delgado, A. Hagerman, A., Schoenbaum, M. (2015). Epidemiologic and economic impacts of applying alternative control strategies for FMD in cattle feedlot operations, International Society for Veterinary Epidemiology and Economics; Merida, Mexico, November. FLAPS (2013). Farm Location and Animal Population Simulator, v1.0. Colorado State University. Software available at Gibbens, J.C., Sharpe, C.E., Wilesmith, J.W., Mansley, L.M., Michalopoulou, E., Ryan, J.B.M., Hudson, M. (2001). Descriptive epidemiology of the 2001 foot-and-mouth disease epidemic in Great Britain: The first five months. Vet Record, 149 (24), Harvey, N., Reeves, A., Schoenbaum, M.A., Zagmutt-Vergara, F.J., Dubé, C., Hill, A.E., Corso, B.A., McNab, W.B., Cartwright, C.I., Salman, M.D. (2007). The North American animal disease spread model: A simulation model to assist decision making in evaluating animal disease incursions. Prev Vet Med 82, McLaws, M., Ribble, C. (2007). Description of recent foot and mouth disease outbreaks in nonendemic areas: Exploring the relationship between detection and epidemic size. Can Vet J, 48 (10), McReynolds, S.W., Sanderson, M.W., Reeves, A., Hill, A.E. (2014). Modeling the impact of vaccination control strategies on a foot and mouth disease outbreak in the Central United States. Prev Vet Med, 117(3-4), Miller, G. Gale, S.B., Eshelman, C.E., Wells, S.J. (2015). Emergency vaccination use in a modelled foot and mouth disease outbreak in Minnesota. Rev Sci Tech, 34(3), Paarlberg, P.L., Hillberg Seitzinger, A., Lee, J.G., Matthews, K.H. (2008). Economic impacts of foreign animal disease (ERR-57). Washington, DC: U.S. Department of Agriculture, Economic Research Service. Pendell, D.L., March, T.L., Coble, K.H., Lusk, J.L., Szmania, S.C. (2015). Economic assessment of FMDv release from the National Bio and Agro Defense Facility. Plos One, 10(6), e Thompson, D. Muriel, P., Russell, D., Osborne, P., Bromley, A., Rowland, M., Creigh-Tyte, S., Brown, C. (2002). Economic costs of the foot and mouth disease outbreak in the United Kingdom in Rev Sci Tech 21(3), Ward, M.P., Highfield, L.D., Vongseng, P., Garner, G.M. (2009). Simulation of foot-and-mouth disease spread within an integrated livestock system in Texas, USA. Prev Vet Med, 88(4), Paper No. 9 Page 10 of 10